NVIDIA releases Nemotron 3.5 Embed 8B, scoring 61.17 nDCG@10 on ViDoRe v3.1 among dense retrievers

NVIDIA has released Nemotron 3.5 Embed, Rerank and Parse, three open, commercially available retrieval models. The 8B embedding model scores 61.17 nDCG@10 on ViDoRe v3.1, which NVIDIA says ranks first among dense retrieval models on that leaderboard as of October 20, 2026.

What shipped

The release adds five checkpoints to the Nemotron Retriever collection:

Model	Role
Nemotron-3.5-Embed-8B-BF16	Text and image retrieval, multilingual; quality-focused
Nemotron-3.5-Embed-1B-BF16	1B model distilled from the 8B, for a smaller footprint
Nemotron-3.5-Embed-1B-NVFP4	NVFP4-quantized 1B; text and image encoders both quantized
Nemotron-3.5-Rerank-8B-BF16	8B cross-encoder scoring query-document pairs across text and images
Nemotron-3.5-Parse (631M parameters)	Document extraction

The embedding models extend the earlier Nemotron 3 Embed text models to document images, including tables, charts, infographics, datasheets and schematics. The source material available for this article does not cover Nemotron 3.5 Parse in detail. Context window, license terms, training cutoff and hosted pricing were not disclosed in the material reviewed.

Benchmark results (NVIDIA-reported)

• ViDoRe v3.1: The Embed 8B scores 61.17 nDCG@10, up 8.94 points from the 52.23 of the previous Llama Nemotron Embed VL 1B.
• 1B distillation: The Embed 1B BF16 scores 58.38, which is 95.4% of the 8B teacher's score.
• MIRACL Vision (18 languages): The 8B scores 72.6 nDCG@10, 4.3 points above Qwen3-VL-Embedding-8B.
• KoViDoRe: NVIDIA says both the 8B and 1B models beat the Qwen baselines on combined image-and-text inputs. No scores were given in the reviewed text.
• Text retrieval: Evaluated across 34 languages.

NVIDIA also compared the models against EmbeddingGemma-2 and Qwen3-VL-Embedding. The leaderboard ranking and comparisons come from NVIDIA's own evaluation.

Agentic retrieval claims

NVIDIA tested retrieval configurations on ViDoRe v3, BRIGHT and BrowseComp-Plus. The agent LLM, tools, task set, corpus and execution policy stayed fixed, and only the search tool's retrieval stack changed.

• Adding Nemotron 3.5 Rerank cut total agent token cost per query by about 14% with Embed 8B and 13% with Embed 1B. NVIDIA says retrieval accuracy also improved in both setups.
• With Embed 8B as the retriever, Nemotron 3.5 Rerank scored about 0.9 nDCG@10 points above Qwen3-VL-Reranker-8B at a similar agentic token cost.

NVIDIA does not name the agent LLM in the reviewed text, so the token savings cannot be generalized to other agents.

Domain adaptation

NVIDIA's fine-tuning recipes produced relative nDCG@10 gains of 18% for Embed 1B BF16 and 3% for Embed 8B BF16 in its chart and infographic experiments. NVIDIA also launched a ViDoRe v3.1 reranking leaderboard with the MTEB team, covering image-only, text-only and combined inputs.

What this means

The notable result is the 1B model, which keeps 95.4% of the 8B's ViDoRe v3.1 score. An NVFP4 variant of it makes visual document retrieval cheaper to index at scale, though NVFP4 needs supported hardware, and no quantized-model scores were published in the reviewed text. The reranker claim is also a cost argument: NVIDIA says better evidence cuts agent token use, and a 13-14% saving matters at enterprise volumes. Because NVIDIA ran all of these comparisons, independent replication on the new MTEB reranking leaderboard will show whether the 0.9-point margin over Qwen3-VL-Reranker-8B holds.